8/1/2017

1

Predictive Data Analytics

Ivo D. DinovStatistics Online Computational ResourceHealth Behavior and Biological Sciences

Michigan Institute for Data Science

University of Michigan

www.SOCR.umich.edu

The University of Michigan (est. 1817)

1917 Winter Spring Summer Fall

8/1/2017

2

Outline

Driving biomedical & health challenges

Common characteristics of Big Data

Data science & predictive analytics

Case-studies

Applications to Neurodegenerative Disease

Data Dashboarding

Compressive Big Data Analytics (CBDA)

Tomorrow’s Healthcare: The Age of Disruptions

Demo(s)

Driving Biomedical/Health Challenges

http://DSPA.predictive.space

Neurodegeneration: Structural Neuroimaging in Alzheimer’s Disease illustrates the Big Data challenges in modeling complex neuroscientific data. 808 ADNI subjects, 3 groups: 200 subjects with Alzheimer’s disease (AD), 383 subjects with mild cognitive impairment (MCI), and 225 asymptomatic normal controls (NC). The 80 neuroimaging biomarkers and 80 highly-associated SNPs.

8/1/2017

3

Driving biomedical/health challenges Phenotype-Genotype-Environmental

Phenotype:Outward manifestations of the form, shape, or characteristics of specific individuals or cohorts

Nature/Genome:Read the genome protein synthesis 2 basic functions

structural proteins determine physical traits or functional traits of protein enzymes catalyzing chemical reactions

Nurture/Environment:Ambient environment directly effects Pheno + Geno Provides raw materials needed for the synthetic processes controlled by Nature/Genes Organisms either synthesize of obtain amino acids with their diet

Cha

nce

SN

R=

Characteristics of Big Biomed Data

Dinov, et al. (2016) PMID:26918190 

Example: analyzing observational data of 1,000’s Parkinson’s disease patients based on 10,000’s signature biomarkers derived from multi-source imaging, genetics, clinical, physiologic, phenomics and demographic data elements.

Software developments, student training, service platforms and methodological advances associated with the Big Data Discovery Science all present existing opportunities for learners, educators, researchers, practitioners and policy makers

IBM Big Data 4V’s: Volume, Variety, Velocity & Veracity

Big Bio Data Dimensions

Tools

SizeHarvesting and management of vast amounts of data

ComplexityWranglers for dealing with heterogeneous data

IncongruencyTools for data harmonization and aggregation

Multi-sourceTransfer and joint modeling of disparate elements

Multi-scaleMacro to meso to micro scale observations

IncompleteReliable management of missing data

8/1/2017

4

Data science & predictive analytics Data science: an emerging extremely transdisciplinary field -

bridging between the theoretical, computational, experimental, and biosocial areas. Deals with enormous amounts of complex, incongruent and dynamic data from multiple sources. Aims to develop algorithms, methods, tools and services capable of ingesting such datasets and generating semi-automated decision support systems

Predictive analytics: utilizing advanced mathematical formulations, powerful statistical computing algorithms, efficient software tools and services to represent, interrogate and interpret complex data. Aims to forecast trends, predict patterns in the data, or prognosticate the process behavior either within the range or outside the range of the observed data (e.g., in the future, or at locations where data may not be available)

http://DSPA.predictive.space

BD

Big Data Information Knowledge ActionRaw Observations Processed Data Maps, Models Actionable Decisions

Data Aggregation Data Fusion Causal Inference  Treatment Regimens

Data Scrubbing Summary Stats Networks, Analytics Forecasts, Predictions

Semantic‐Mapping Derived Biomarkers Linkages, Associations Healthcare Outcomes

I K A

Dinov, et al. (2016) PMID:26918190 

8/1/2017

5

Case‐Studies – ALS

Data Source

Sample Size/Data Type Summary

ProActArchive

Over 100 variables are recorded for all subjects including: Demographics: age, race, medical history, sex; Clinical data: Amyotrophic Lateral Sclerosis Functional Rating Scale (ALSFRS), adverse events, onset_delta, onset_site, drugs use (riluzole) The PRO-ACT training dataset contains clinical and lab test information of 8,635 patients. Information of 2,424 study subjects with valid gold standard ALSFRS slopes will be used in out processing, modeling and analysis

The time points for all longitudinally varying data elements will be aggregated into signature vectors. This will facilitate the modeling and prediction of ALSFRS slope changes over the first three months (baseline to month 3)

Identify highly predictive classifiers to detect, track and prognosticate the progression of ALS (in terms of clinical outcomes like ALSFRS and muscle function)

Provide a decision tree prediction of adverse events based on subject phenotype and 0-3 month clinical assessment changes

Case‐Studies – Parkinson’s Disease  Predict the clinical diagnosis of patients using all available data (with and

without the UPDRS clinical assessment, which is the basis of the clinical diagnosis by a physician)

Compute derived neuroimaging and genetics biomarkers that can be used to model the disease progression and provide automated clinical decisions support

Generate decision trees for numeric and categorical responses (representing clinically relevant outcome variables) that can be used to suggest an appropriate course of treatment for specific clinical phenotypes

Data Source Sample Size/Data Type Summary

PPMIArchive

Demographics: age, medical history, sex.Clinical data: physical, verbal learning and language, neurological and olfactory, UPSIT, UPDRS scores, ADL, GDS-15, …Imaging data: structural MRI.Genetics data: APOE genotypes e2/e3Cohorts: Group 1 = {PD Subjects}, N1 = 263; Group 2 = {PD Subjects with Scans without Evidence of a Dopaminergic Deficit (SWEDD)}, N2 = 40; Group 3 = {Control Subjects}, N3 = 127.

The longitudinal PPMI dataset including clinical, biological and imaging data (screening, baseline, 12, 24, and 48 month follow-ups) may be used conduct model-based predictions as well as model-free classification and forecasting analyses

8/1/2017

6

2 20005 Ongoing characteristics Email access2 110007 Ongoing characteristics Newsletter communications, date sent100 25780 Brain MRI Acquisition protocol phase.100 12139 Brain MRI Believed safe to perform brain MRI scan100 12188 Brain MRI Brain MRI measurement completed100 12187 Brain MRI Brain MRI measuring method100 12663 Brain MRI Reason believed unsafe to perform brain MRI100 12704 Brain MRI Reason brain MRI not completed100 12652 Brain MRI Reason brain MRI not performed101 12292 Carotid ultrasound Carotid ultrasound measurement completed101 12291 Carotid ultrasound Carotid ultrasound measuring method101 20235 Carotid ultrasound Carotid ultrasound results package101 22672 Carotid ultrasound Maximum carotid IMT (intima‐medial thickness) at 120 degrees 101 22675 Carotid ultrasound Maximum carotid IMT (intima‐medial thickness) at 150 degrees 101 22678 Carotid ultrasound Maximum carotid IMT (intima‐medial thickness) at 210 degrees 101 22681 Carotid ultrasound Maximum carotid IMT (intima‐medial thickness) at 240 degrees 101 22671 Carotid ultrasound Mean carotid IMT (intima‐medial thickness) at 120 degrees 101 22674 Carotid ultrasound Mean carotid IMT (intima‐medial thickness) at 150 degrees 101 22677 Carotid ultrasound Mean carotid IMT (intima‐medial thickness) at 210 degrees 101 22680 Carotid ultrasound Mean carotid IMT (intima‐medial thickness) at 240 degrees 101 22670 Carotid ultrasound Minimum carotid IMT (intima‐medial thickness) at 120 degrees 101 22673 Carotid ultrasound Minimum carotid IMT (intima‐medial thickness) at 150 degrees 101 22676 Carotid ultrasound Minimum carotid IMT (intima‐medial thickness) at 210 degrees 101 22679 Carotid ultrasound Minimum carotid IMT (intima‐medial thickness) at 240 degrees 101 22682 Carotid ultrasound Quality control indicator for IMT at 120 degrees101 22683 Carotid ultrasound Quality control indicator for IMT at 150 degrees101 22684 Carotid ultrasound Quality control indicator for IMT at 210 degrees

Case‐Studies – General Populations

UK Biobank – discriminate between HC, single and multiple comorbid conditions

Predict likelihoods of various developmental or aging disorders

Forecast cancer

Data Source Sample Size/Data Type Summary

UK Biobank

Demographics: > 500K casesClinical data: > 4K featuresImaging data: T1, resting‐state fMRI, task fMRI, T2_FLAIR, dMRI, SWI Genetics data

The longitudinal archive ofentire UK population (NHS)

Integrated & Transdisciplinary

‐ Research‐ Education‐ Practice

8/1/2017

7

Big Data Analytics Resourceome

http://socr.umich.edu/docs/BD2K/BigDataResourceome.html

End‐to‐end Pipeline Workflow Solutions 

Dinov, et al., 2014, Front. Neuroinform.;                               Dinov, et al., Brain Imaging & Behavior, 2013

8/1/2017

8

Predictive Big Data Analytics in Parkinson’s Disease Big Data: Parkinson’s Progression Markers Initiative (PPMI). Defining data characteristics –

large size, incongruency, incompleteness, complexity, multiplicity of scales, and heterogeneity of sources (imaging, genetics, clinical, demographic)

Approach – machine-learning based classification – introduce methods for rebalancing imbalanced cohorts, – utilize a wide spectrum of classification methods to generate phenotypic predictions, – reproducible machine-learning based classification

Results- Predicted Parkinson’s disease in the PPMI subjects (consistent accuracy, sensitivity, and

specificity exceeding 96%- Confirmed using internal statistical 5-fold cross-validation- Clinical features: Unified Parkinson's Disease Rating Scale (UPDRS) scores

demographic (e.g., age), genetics (e.g., rs34637584, chr12) - Neuroimaging biomarkers (e.g., cerebellum shape index)

Model-free Big Data machine learning-based classification methods (Adaptive boosting, support vector machines) outperform model-based techniques (GEE, GLM, MEM) in terms of predictive precision and reliability (e.g., forecasting patient diagnosis).

UPDRS scores play a critical role in predicting diagnosis, which is expected based on the clinical definition of Parkinson’s disease.

Excluding longitudinal UPDRS data, the accuracy of model-free machine learning based classification is over 80%. The methods, software and protocols are openly shared and can be employed to study other neurodegenerative disorders

Dinov, et al., (2016) PMID:27494614

Probabilistic Bayesian Inference – Chance Encounters

Suppose a patient visits a primary care clinic and is seen by a male provider not wearing a badge/insignia

To address the clinician appropriately, the patient is trying to figure out if he is more likely to be a doctor(D) or a nurse (N).

Notation , M M , , and .

Traditional stereotypes may suggest that a male provider is more likely to be a doctor than a nurse.

Is the odds likelihood ratio, |

|1?

8/1/2017

9

Probabilistic Bayesian Inference – Chance Encounters

Dinov, et al., (2016) PMID: 26998309

Actually, the odds are that the Male healthcare provider is a Nurse!

||

|

|||

11323

4,500,000

435,000326

10.3 1.2.

An odds likelihood ratio >1 dispels an initial stereotypic vision of females and males as predominantly nurses and physicians, respectively.

This is a simple example of data-driven/evidence-based Inference.

Dat

a: K

aise

r F

amily

Fou

ndat

ion

Predictive Big Data Analytics:Applications to Parkinson’s Disease

Varplot

• Critical predictive data elements (Y-axis)

• Their impact scores (X-axis)

AdaBoost classifier for Controls vs. Patients prediction

MLclassifier

accuracy sensitivity specificitypositive 

predictive valuenegative 

predictive valuelog odds ratio 

(LOR)

AdaBoost 0.996324 0.994141 0.998264 0.9980392 0.9948097 11.4882058

SVM 0.985294 0.994140 0.977431 0.9750958 0.9946996 8.902166

Dinov, et al., (2016) PMID:27494614

8/1/2017

10

Tools Developed, Validated & Shared

End-to-end Big Data analytic protocol jointly processing complex imaging, genetics, clinical, demo data for assessing PD risk

o Methods for rebalancing of imbalanced cohorts

o ML classification methods generating consistent and powerful phenotypic predictions

o Reproducible protocols for extraction of derived neuroimaging and genomics biomarkers for diagnostic forecasting

19https://github.com/SOCR/PBDA

SOCR Big Data Dashboardhttp://socr.umich.edu/HTML5/Dashboard

Web‐service combining and integrating multi‐source socioeconomic and medical datasets 

Big data analytic processing

Interface for exploratory navigation, manipulation and visualization

Adding/removing of visual queries and interactive exploration of multivariate associations

Powerful HTML5 technology enabling mobile on‐demand computing

Husain, et al., 2015, PMID:26236573

8/1/2017

11

SOCR Dashboard (Exploratory Big Data Analytics): Data Fusion

http://socr.umich.edu/HTML5/Dashboard

SOCR Dashboard (Exploratory Big Data Analytics): Associations

8/1/2017

12

Compressive Big Data Analytics (CBDA)

o Foundation for Compressive Big Data Analytics (CBDA)

o Iteratively generate random (sub)samples from the Big Data collection

o Then, using classical techniques to obtain model-based or non-parametric inference based on the sample

o Next, compute likelihood estimates (e.g., probability values quantifying effects, relations, sizes)

o Repeat – the process continues iteratively until a criterion is met– the (re)sampling and inference steps many times (with or without using the results of previous iterations as priors for subsequent steps)

Dinov, 2016, PMID:  26998309 

CBDA Framework

8/1/2017

13

FAIR Data & Open‐Science Principles

Share resources

Collaborate

Permissive licenses (e.g., LGPL/CC-BY)

Project management (e.g., GitHub/Jira)

Open-access pubs

Public-private partnerships

Co-mentoring of trainees

Effective transdisciplinary methods

Resource Interoperability

Result Reproducibility

Tomorrow’s Healthcare:  The Age of Disruptions

Address Some Challenging Open Problems Powerful data wrangling strategies Techniques for data harmonization, appending, aggregation Mathematical framework for Big Data representation (cf. 6D, CBDA) Reliable and secure Biomed/Health data communication/sharing Advanced machine-learning decision support systems

Future Healthcare Innovation & Delivery On-demand, service-oriented, geo-location-agnostic health delivery Rapid deployment, continuous development/innovation/refinement (Evidence-based) Data Science and Predictive Health Analytics Personalized Medicine (from diagnosis, to treatment and prognosis) End-to-end Doctronic (Human-Machine) services – Clinical Decision

Support Systems (improve overall population health, reduce costs, better prognostication, enhanced reliability, rapid response)

Some examples …

8/1/2017

14

Clinical Decision Support Doctronics

Hospital Admissions Survival Inference and Clinical outcome forecasting using a hospital

admissions (N=58,863 and k=9): Admission_Length: Duration of hospital stay (in days)

Death: Indicator of Death (1) or survival (0)

…

Demographics & (Human-labeled) Diagnoses

VarImpPlot:    RF Survival Prediction

hosAdmissions_RFmodel_pred 0    10 9937  9631  722  151                 

Accuracy : 0.8568759             95% CI : (0.85, 0.86)            

Sensitivity : 0.9322638             Specificity : 0.1355476             Pos Pred Value : 0.9116514             Neg Pred Value : 0.1729668             

NeuralNet Deep Learning: Survival Probability (N=11,773)

Test/Validation Cases

Probab

ility

Clinical Decision Support Doctronics

Top 1 Population characteristics1 2 Ongoing characteristics100003 100 Brain MRI100006 101 Carotid ultrasound100003 102 Heart MRI100003 103 DXA assessment100006 104 ECG at rest, 12-lead100003 105 Abdominal MRI100 106 Task functional brain MRI100 107 Diffusion brain MRI100 108 Scout images and configuration for brain MRI100 109 Susceptibility weighted brain MRI100 110 T1 structural brain MRI100 111 Resting functional brain MRI100 112 T2-weighted brain MRI100088 113 Local environment113 114 Residential air pollution113 115 Residential noise pollution100088 116 Cognitive function follow-up116 117 Pairs matching test116 118 Fluid intelligence test116 120 Numeric memory test116 121 Trail making test116 122 Symbol digit substitution test100088 123 Work environment103 124 Body composition by DXA103 125 Bone size, mineral and density by DXA

Population-wide Study of Health and DiseaseNational Health System

Longitudinal Data (N>1M, k>4,300)

DoB

Computable Phenotypes – Genomics, Imaging, Biospecimen, Clinical, Environmental

GWASHumanConnectome

8/1/2017

15

Clinical Decision Support Doctronics

Personalized medicine – Traumatic Brain Injury (TBI)

LONI/USC, BIRC/UCLA, SOCR/UMich

Acknowledgments 

FundingNIH: P20 NR015331, U54 EB020406, P50 NS091856, P30 DK089503, P30AG053760

NSF: 1636840, 1416953, 0716055, 1023115

The Elsie Andresen Fiske Research Fund

Collaborators • SOCR: Alexandr Kalinin, Selvam Palanimalai, Syed Husain, Matt Leventhal, Ashwini Khare, Rami Elkest,

Abhishek Chowdhury, Patrick Tan, Gary Chan, Andy Foglia, Pratyush Pati, Brian Zhang, Juana Sanchez, Dennis Pearl, Kyle Siegrist, Rob Gould, Jingshu Xu, Nellie Ponarul, Ming Tang, Asiyah Lin, Nicolas Christou, Hanbo Sun, Tuo Wang

• LONI/INI: Arthur Toga, Roger Woods, Jack Van Horn, Zhuowen Tu, Yonggang Shi, David Shattuck, Elizabeth Sowell, Katherine Narr, Anand Joshi, Shantanu Joshi, Paul Thompson, Luminita Vese, Stan Osher, Stefano Soatto, Seok Moon, Junning Li, Young Sung, Carl Kesselman, Fabio Macciardi, Federica Torri

• UMich MIDAS/MNORC/AD/PD Centers: Cathie Spino, Chuck Burant, Ben Hampstead,

Stephen Goutman, Stephen Strobbe, Hiroko Dodge, Hank Paulson, Bill Dauer

http://SOCR.umich.edu